DeepMind AI

[Pondera]

[html]https://www.reddit.com/r/programming/co ... _the_best/[/html]

Interesting link to new Artificial Intelligence applications. There's a paper there which goes into depth and the youtube links to a couple of actual games is interesting too. What's mind blowing is that this program learned Chess based on the rules alone and then competed against itself for four hours, only to become a better chess engine than the best out there.

I'd be interested to see if this program could be used for linguistics. Ie. teach it the rules of grammar and give it a dictionary - then let it converse with itself for a few days!

[chownah]

[html]https://www.reddit.com/r/programming/co ... _the_best/[/html]

Interesting link to new Artificial Intelligence applications. There's a paper there which goes into depth and the youtube links to a couple of actual games is interesting too. What's mind blowing is that this program learned Chess based on the rules alone and then competed against itself for four hours, only to become a better chess engine than the best out there.

I'd be interested to see if this program could be used for linguistics. Ie. teach it the rules of grammar and give it a dictionary - then let it converse with itself for a few days!

I'm only guessing here....just my own ideas based on not much.
I think that while some aspects of this might apply to language it probably in itself will not be the answer for languages. The difference I see between games and language is that in language there is not a precisely defined idea of what winning is.....the goal in language is to communicate ideas between entities. This learning algorithm might be able to create the worlds most efficient language by talking to itself but then no one and no computer would know the language. (A side note: already some AI systems are creating their own language which they use to internally represent concepts of their own creation (from what I gather) and it sort of scares researchers in that they can not understand how the AI system achieves its results....with the results being demonstrably valid.) I'm not saying that AI can't progress to handle natural languages but I'm just saying that I don't think that this approach is the one which will do the job.
I'm only guessing here....just my own ideas based on not much.
chownah

[Pondera]

[html]https://www.reddit.com/r/programming/co ... _the_best/[/html]

Interesting link to new Artificial Intelligence applications. There's a paper there which goes into depth and the youtube links to a couple of actual games is interesting too. What's mind blowing is that this program learned Chess based on the rules alone and then competed against itself for four hours, only to become a better chess engine than the best out there.

I'd be interested to see if this program could be used for linguistics. Ie. teach it the rules of grammar and give it a dictionary - then let it converse with itself for a few days!

I'm only guessing here....just my own ideas based on not much.
I think that while some aspects of this might apply to language it probably in itself will not be the answer for languages. The difference I see between games and language is that in language there is not a precisely defined idea of what winning is.....the goal in language is to communicate ideas between entities. This learning algorithm might be able to create the worlds most efficient language by talking to itself but then no one and no computer would know the language. (A side note: already some AI systems are creating their own language which they use to internally represent concepts of their own creation (from what I gather) and it sort of scares researchers in that they can not understand how the AI system achieves its results....with the results being demonstrably valid.) I'm not saying that AI can't progress to handle natural languages but I'm just saying that I don't think that this approach is the one which will do the job.
I'm only guessing here....just my own ideas based on not much.
chownah

No. Definitely. The purpose of conversing would have to be well defined for the algorithm to achieve any usefulness. Just carrying on a conversation for half an hour might be a purpose. Psychotherapy might be the goal. Helping a person understand their inhibitions and how to overcome them.

[Sam Vara]

I'd be interested to see if this program could be used for linguistics. Ie. teach it the rules of grammar and give it a dictionary - then let it converse with itself for a few days!

The results might well be incomprehensible. As Wittgenstein said, "If a lion could speak, we should not be able to understand him".

[Pondera]

I'd be interested to see if this program could be used for linguistics. Ie. teach it the rules of grammar and give it a dictionary - then let it converse with itself for a few days!

The results might well be incomprehensible. As Wittgenstein said, "If a lion could speak, we should not be able to understand him".

True. The lion would speak of "lion" things. But I see no reason why we would have a different concept of, say, a giraffe - for instance. In which case it would be a matter of figuring out an efficient way of translating words.

The program is another story. If it can engage with the world it might be able to form concepts. The importance of an object in the world might be translatable. In that case all we would need to do is teach it to name the concepts it learns. But i think chownah already alluded to this - programs conceptualize in different ways. The translations are all idioms.